1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
|
// Copyright (C) 2018-2019 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//
#include "ext_list.hpp"
#include "ext_base.hpp"
#include <cmath>
#include <string>
#include <vector>
#include <cassert>
#include "ie_parallel.hpp"
namespace InferenceEngine {
namespace Extensions {
namespace Cpu {
class PadImpl: public ExtLayerBase {
public:
explicit PadImpl(const CNNLayer* layer) {
try {
if (layer->insData.empty() || layer->outData.empty())
THROW_IE_EXCEPTION << layer->name << " Incorrect number of input/output edges!";
pads_begin = layer->GetParamAsUInts("pads_begin");
std::vector<unsigned int> pads_end = layer->GetParamAsUInts("pads_end");
src_dims = layer->insData[0].lock()->getTensorDesc().getDims();
dst_dims = layer->outData[0]->getTensorDesc().getDims();
if (src_dims.size() != dst_dims.size() || pads_begin.size() != src_dims.size())
THROW_IE_EXCEPTION << layer->name << " Incorrect number of input/output dimensions!";
std::string pad_mode = layer->GetParamAsString("pad_mode");
if (pad_mode == "constant") {
padMode = CONSTANT;
} else if (pad_mode == "edge") {
padMode = EDGE;
} else if (pad_mode == "reflect") {
padMode = REFLECT;
for (size_t i = 0; i < src_dims.size(); i++) {
if ((src_dims[i] - 1) < pads_begin[i] || (src_dims[i] - 1) < pads_end[i])
THROW_IE_EXCEPTION << layer->name << " Incorrect pads_begin or pads_end for 'reflect' pad mode";
}
} else if (pad_mode == "symmetric") {
padMode = SYMMETRIC;
for (size_t i = 0; i < src_dims.size(); i++) {
if (src_dims[i] < pads_begin[i] || src_dims[i] < pads_end[i])
THROW_IE_EXCEPTION << layer->name << " Incorrect pads_begin or pads_end for 'symmetric' pad mode";
}
} else {
THROW_IE_EXCEPTION << layer->name
<< " Incorrect pad_mode. Only constants|edge|reflect|symmetric modes are supported!";
}
if (padMode == CONSTANT)
pad_value = layer->GetParamAsFloat("pad_value", 0.f);
srcStrides = layer->insData[0].lock()->getTensorDesc().getBlockingDesc().getStrides();
dstStrides = layer->outData[0]->getTensorDesc().getBlockingDesc().getStrides();
work_amount = dst_dims[0] * dstStrides[0];
for (size_t i = 0; i < src_dims.size(); i++)
src_o_dms.push_back(src_dims[i] + pads_begin[i]);
addConfig(layer, { DataConfigurator(ConfLayout::PLN) }, { DataConfigurator(ConfLayout::PLN) });
} catch (InferenceEngine::details::InferenceEngineException &ex) {
errorMsg = ex.what();
}
}
StatusCode execute(std::vector<Blob::Ptr>& inputs, std::vector<Blob::Ptr>& outputs, ResponseDesc *resp) noexcept override {
const float *src_data = inputs[0]->cbuffer().as<const float *>() +
inputs[0]->getTensorDesc().getBlockingDesc().getOffsetPadding();
float* dst_data = outputs[0]->cbuffer().as<float *>() +
outputs[0]->getTensorDesc().getBlockingDesc().getOffsetPadding();
switch (padMode) {
case CONSTANT:
pad_constant(src_data, dst_data);
break;
case EDGE:
pad_edge(src_data, dst_data);
break;
case REFLECT:
pad_reflect(src_data, dst_data);
break;
case SYMMETRIC:
pad_symmetric(src_data, dst_data);
break;
default:
return GENERAL_ERROR;
}
return OK;
}
private:
enum PadMode {
CONSTANT = 0,
EDGE = 1,
REFLECT = 2,
SYMMETRIC = 3
};
void pad_constant(const float *src_data, float* dst_data);
void pad_edge(const float *src_data, float* dst_data);
void pad_reflect(const float *src_data, float* dst_data);
void pad_symmetric(const float *src_data, float* dst_data);
PadMode padMode = CONSTANT;
float pad_value = 0.f;
SizeVector src_dims;
SizeVector dst_dims;
std::vector<unsigned int> pads_begin;
SizeVector src_o_dms;
SizeVector srcStrides;
SizeVector dstStrides;
size_t work_amount;
};
inline size_t parallel_init(size_t start, size_t size, std::vector<size_t> &counters, std::vector<size_t> &dims) {
for (int j = size - 1; j >= 0; j--) {
counters[j] = start % dims[j];
start = start / dims[j];
}
return start;
}
inline void parallel_step(size_t size, std::vector<size_t> &counters, std::vector<size_t> &dims) {
for (int j = size - 1; j >= 0; j--) {
counters[j] = (counters[j] + 1) % dims[j];
if (counters[j] != 0)
return;
}
}
void PadImpl::pad_constant(const float *src_data, float* dst_data) {
int offset = 0;
for (size_t i = 0; i < srcStrides.size(); ++i)
offset += pads_begin[i] * srcStrides[i];
parallel_nt(0, [&](const int ithr, const int nthr) {
size_t start = 0, end = 0;
SizeVector counters(dst_dims.size(), 0);
splitter(work_amount, nthr, ithr, start, end);
parallel_init(start, dst_dims.size(), counters, dst_dims);
for (size_t iwork = start; iwork < end; ++iwork) {
int srcIdx = 1;
int dstIdx = 0;
for (size_t i = 0; i < dstStrides.size(); ++i)
dstIdx += counters[i] * dstStrides[i];
for (size_t i = 0; i < counters.size(); ++i) {
if (counters[i] < pads_begin[i] || counters[i] >= src_o_dms[i]) {
dst_data[dstIdx] = pad_value;
srcIdx = 0;
break;
}
}
if (srcIdx) {
int srcIdx = 0;
for (size_t i = 0; i < srcStrides.size(); ++i)
srcIdx += counters[i] * srcStrides[i];
dst_data[dstIdx] = src_data[srcIdx - offset];
}
parallel_step(dst_dims.size(), counters, dst_dims);
}
});
}
void PadImpl::pad_edge(const float *src_data, float* dst_data) {
parallel_nt(0, [&](const int ithr, const int nthr) {
size_t start = 0, end = 0;
SizeVector counters(dst_dims.size(), 0);
splitter(work_amount, nthr, ithr, start, end);
parallel_init(start, dst_dims.size(), counters, dst_dims);
for (size_t iwork = start; iwork < end; ++iwork) {
int srcIdx = 0;
int dstIdx = 0;
for (size_t i = 0; i < dstStrides.size(); ++i)
dstIdx += counters[i] * dstStrides[i];
for (size_t i = 0; i < srcStrides.size(); ++i) {
int idx = (counters[i] < pads_begin[i]) ? 0 :
((counters[i] >= src_o_dms[i]) ? (src_dims[i] - 1) : (counters[i] - pads_begin[i]));
srcIdx += idx * srcStrides[i];
}
dst_data[dstIdx] = src_data[srcIdx];
parallel_step(dst_dims.size(), counters, dst_dims);
}
});
}
void PadImpl::pad_reflect(const float *src_data, float* dst_data) {
SizeVector src_2;
for (size_t i = 0; i < src_dims.size(); i++)
src_2.push_back(src_dims[i] + src_o_dms[i] - 2);
parallel_nt(0, [&](const int ithr, const int nthr) {
size_t start = 0, end = 0;
SizeVector counters(dst_dims.size(), 0);
splitter(work_amount, nthr, ithr, start, end);
parallel_init(start, dst_dims.size(), counters, dst_dims);
for (size_t iwork = start; iwork < end; ++iwork) {
int srcIdx = 0;
int dstIdx = 0;
for (size_t i = 0; i < dstStrides.size(); ++i)
dstIdx += counters[i] * dstStrides[i];
for (size_t i = 0; i < srcStrides.size(); ++i) {
int idx = (counters[i] < pads_begin[i]) ? (pads_begin[i] - counters[i]) :
((counters[i] >= src_o_dms[i]) ? (src_2[i] - counters[i]) : (counters[i] - pads_begin[i]));
srcIdx += idx * srcStrides[i];
}
dst_data[dstIdx] = src_data[srcIdx];
parallel_step(dst_dims.size(), counters, dst_dims);
}
});
}
void PadImpl::pad_symmetric(const float *src_data, float* dst_data) {
SizeVector src_2;
for (size_t i = 0; i < src_dims.size(); i++)
src_2.push_back(src_dims[i] + src_o_dms[i] - 1);
parallel_nt(0, [&](const int ithr, const int nthr) {
size_t start = 0, end = 0;
SizeVector counters(dst_dims.size(), 0);
splitter(work_amount, nthr, ithr, start, end);
parallel_init(start, dst_dims.size(), counters, dst_dims);
for (size_t iwork = start; iwork < end; ++iwork) {
int srcIdx = 0;
int dstIdx = 0;
for (size_t i = 0; i < dstStrides.size(); ++i)
dstIdx += counters[i] * dstStrides[i];
for (size_t i = 0; i < srcStrides.size(); ++i) {
int idx = (counters[i] < pads_begin[i]) ? (pads_begin[i] - 1 - counters[i]) :
((counters[i] >= src_o_dms[i]) ? (src_2[i] - counters[i]) : (counters[i] - pads_begin[i]));
srcIdx += idx * srcStrides[i];
}
dst_data[dstIdx] = src_data[srcIdx];
parallel_step(dst_dims.size(), counters, dst_dims);
}
});
}
REG_FACTORY_FOR(ImplFactory<PadImpl>, Pad);
} // namespace Cpu
} // namespace Extensions
} // namespace InferenceEngine
|